Performance of selective speech features for speaker identification

Different types of speech features are used for speaker identification. Speeeh features sueh as Mel'frequency cepstral coefficients (MFGG), Jog area ratios (LAW, arcsin reflection coefficients (ARC), cepstralcoefficients (CO) and reflection coefficients (RE)are shown to produce promising results for speaker identification, The results produced are with various speaker databases and with different classifiers. In this work, the relative performance of these selective speech features has been evaluated with a single speaker database. A novel statistical measure, feature discriminant measure (FDfflhas been proposed to evaluate the relative perfermance of these selective features. Different statistical properties, such as probability density characteristics, K'S test and F-ratio are also used to evaluate the performances of these speech features. For comparison, a vector quantisation (VQ)classifier is used. The MFGG and LAR features show the best results with 10096 speaker identification. The ARG features, the GG features and RG features show 99.5%, 99% and 97% speaker identification results respectively with the VQ classifier. These relative performances mateh with the values of the proposed feature discriminant measures.